JP4427074B2 - Plant control equipment - Google Patents

Description

  The present invention relates to a plant control device, and more particularly to a thermal power plant control device.

  In the plant control device, the measurement signal of the operating state quantity obtained from the plant to be controlled is processed, and the operation signal to be given to the plant to be controlled is calculated and transmitted as a control command to the control device.

  In such a plant control apparatus, an algorithm for calculating an operation signal is mounted so that the measurement signal of the operation state quantity of the plant satisfies the target value.

  As a control algorithm used for plant control, there is a PI (proportional / integral) control algorithm. In PI control, a value obtained by integrating the deviation with time is added to a value obtained by multiplying the deviation between the measurement signal of the plant and its target value by a proportional gain to derive an operation signal to be given to the controlled object.

  On the other hand, since the control algorithm is automatically corrected / changed in accordance with changes in the operation state and environment of the plant, the plant can also be controlled using adaptive control or a learning algorithm.

  As a method of deriving an operation signal of a control device that controls a plant using a learning algorithm, Japanese Patent Application Laid-Open No. 2000-35956 describes a technique related to a control device using a reinforcement learning method using a model.

  In the method based on the technique using the reinforcement learning method, the control device has a model that predicts the characteristics of the control target and a learning unit that learns a model input generation method that achieves the target value of the model output.

  Then, by inputting the model input learned by the learning unit into the model, an effect that the model output approaches the target value can be obtained.

  In such learning-type adaptive control, a control algorithm for controlling the plant with high accuracy is obtained by correcting the model using the measurement signal obtained by measuring the operating state of the plant, and executing learning again using the corrected model. Modify / change online.

  Therefore, in the learning, the model is corrected within a cycle (control cycle) in which the operation signal output from the control device to the plant is changed, the learning is performed again using the corrected model, and the learning is finished. It is necessary to modify / change the control algorithm.

  This control cycle can be regarded as a time until the plant is settled after operating the plant, and generally has a length of several minutes to several tens of minutes.

  For example, in the control of a complex plant represented by a thermal power plant, the number of model input orders is from several tens to several hundreds, and the number of combinations of model inputs to be learned (search space) increases, resulting in a long learning time. It becomes difficult to modify / change the control algorithm necessary for accurately controlling the plant online.

  Therefore, in order to complete the learning within the control cycle, it is necessary to modify the model in response to the increase in the model input order and to speed up the learning that is performed again using the modified model.

  Japanese Patent Laid-Open No. 7-160661 classifies teacher data into a plurality of patterns based on a combination of plant measurement value information in neural network learning, extracts patterns to be learned based on control results, and extracts the extracted patterns. A technique for speeding up learning by learning only teacher data is disclosed.

JP 2000-35956 A Japanese Patent Laid-Open No. 7-160661

  By using the technique of the reinforcement learning method using the model described in Japanese Patent Laid-Open No. 2000-35956, it is possible to automatically learn a method for generating an operation signal that can achieve a control target.

  However, when a control algorithm is modified by correcting the model based on the measurement signal obtained by measuring the operating state of the plant, and then learning again using the corrected model, the model is input at a complex plant. As the order increases, it becomes difficult to perform the learning within the control cycle for controlling the plant.

  Further, when the technique described in Japanese Patent Laid-Open No. 7-160661 is used, it is possible to reduce the search space by dividing the teacher data by patterning, thereby speeding up the learning. That is, even when the model input order increases and the search space (teacher data number) increases, learning within the control period can be performed by appropriately performing patterning.

  However, when the teacher data to be classified is biased, the number of teacher data decreases depending on the pattern to be learned, and a desired learning result cannot be obtained, and a desired learning result cannot be obtained.

  In addition, since the types of patterns to be classified are determined based on the knowledge of the designer of the control device, there is a problem that the burden on manpower becomes very large when generating patterns.

  An object of the present invention is to correct a model based on a measurement signal obtained by measuring the operating state of a plant, and to correct a plant control algorithm by executing learning at a high speed using the corrected model. An object of the present invention is to provide a plant control apparatus for controlling a plant with high accuracy.

The plant control apparatus of the present invention is a plant control apparatus including an operation signal generation unit that calculates an operation signal that becomes a control command for a plant to be controlled using a measurement signal obtained by measuring the operation state of the plant. The control device includes a measurement signal database in which measurement signals of the measured plant are stored, an operation signal database in which operation signals for the plant are stored, a numerical analysis execution unit that analyzes the operation characteristics of the plant, A model that simulates the control characteristics of the plant when an operation signal is given to the plant based on the information of the numerical analysis result from the numerical analysis execution unit, and the model output that is simulated by the model using the model is model output when learning a method of generating a model input so as to achieve the target value, the state input before plant operation starts to be input to the learning section Trained with remains, after the plant operation starts a learning unit that learns by patterning the state input which is input to the learning unit using the pattern data stored in the pattern database, the learning unit pattern The pattern data is selected from the pattern data stored in the database, and the state input is patterned using the selected pattern data to reduce the input order, and the learning result obtained by the learning unit is obtained. a learning information database of training information data is stored that is, the operation signal and the control logic database information used on an operation signal output is stored from the generator, before plant operation initiated by the learning unit I learned learning data and the learning information database on the basis of the stored learned data pattern Day Accuracy and the pattern generator for generating a pattern database in which patterns data generated by the pattern generation unit is stored, the plant if the pattern data used for learning of a plurality from among the learning data of the pattern comprising a learning result judging unit for selecting a better control to effect a good training data, it is configured to calculate the operation signal from the operation signal generation unit based on the learning data selected in the learning result determination unit It is characterized by.

  According to the present invention, the model is corrected based on the measurement signal obtained by measuring the operation state of the plant, and the learning that is learned again using the corrected model is executed at high speed to correct the plant control algorithm. Therefore, a plant control device that controls the plant with high accuracy can be realized.

  Next, a plant control apparatus according to an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 is a control block diagram showing the overall configuration of a plant control apparatus according to an embodiment of the present invention.

  In FIG. 1, the plant according to the present embodiment is a plant 100 that constitutes a thermal power plant including a boiler that uses coal as fuel, and the thermal power plant 100 that includes this boiler is controlled by a control device 200. It is configured.

  The control device 200 that controls the plant 100 to be controlled includes a numerical analysis execution unit 220, a model 230, a learning unit 260, an operation signal generation unit 280, a learning result determination unit 300, and a pattern generation unit 400 as arithmetic devices. It has a provided configuration.

  The control device 200 includes a measurement signal database 210, a numerical analysis database 240, an operation signal database 250, a learning information database 270, a control logic database 290, and a pattern database 500 as databases.

  The control device 200 is provided with an external input interface 201 and an external output interface 202 as interfaces with the outside.

  In the control device 200, a measurement signal 1 obtained by measuring various state quantities of the thermal power plant is taken into the control device 200 from the plant 100, which is a thermal power plant, via the external input interface 201.

  Also, an operation signal 15 for controlling, for example, a boiler burner and an air flow rate of an air port is sent from the control device 200 to the plant 100 that is the thermal power plant to be controlled, via the external output interface 202. .

  In the control device 200, the measurement signal 1 obtained by measuring various state quantities of the plant 100 is stored in the measurement signal database 210 which is a database provided in the control device 200 as the measurement signal 2 after passing through the external input interface 201.

  The operation signal 14 generated by the operation signal generation unit 280 that is an arithmetic device provided in the control device 200 is transmitted to the external output interface 202 and is also input to the operation signal database 250 that is a database provided in the control device 200. Saved.

  The operation signal generation unit 280 outputs the control logic data 13 stored in the control logic database 290 that is a database provided in the control device 200 and the learning result determination unit 300 that is an arithmetic device provided in the control device 200. Using the learned data 12 thus generated, the operation signal 14 to be the operation signal 15 is generated for the plant 100 and output to the external input interface 202 so that the measurement signal 1 of the plant 100 achieves the operation target value. To do.

  The control logic database 290 stores a control circuit for calculating the control logic data 13 and control parameters in order to output the control logic data 13.

  A known PI control can be used for the control circuit for calculating the control logic data 13.

  The learning data stored in the learning information database 270 that is a database provided in the control device 200 is generated by the learning unit 260 that is an arithmetic device provided in the control device 200.

  The learning unit 260 is connected to a model 230 that is an arithmetic device provided in the control device 200.

  The model 230 has a function of simulating the control characteristics of the plant 100 that is a thermal power plant. That is, the model 230 simulates the operation equivalent to giving the operation signal 15 as a control command to the plant 100 and obtaining the measurement signal 1 of the control result.

  For this simulation calculation, the model input 7 for operating the model 230 is received from the learning unit 260, the model 230 simulates the characteristic change due to the control of the plant 100, and the model output 8 of the simulation calculation result is obtained. Configured to get.

  Here, the model output 8 is a predicted value of the measurement signal 1 of the plant 100.

  The model 230 is constructed based on the numerical analysis result 6 stored in the numerical analysis database 240 that is a database provided in the control device 200.

  A numerical analysis execution unit 220 that is an arithmetic device provided in the control device 200 analyzes the operating characteristics of the plant 100 using a physical model that simulates the plant 100.

  The calculation result 4 obtained by analyzing the characteristics of the plant 100 using the physical model in the numerical analysis execution unit 220 is stored in the numerical analysis database 240.

  In the model 230, the model output 8 corresponding to the model input 7 is calculated using a statistical technique such as a neural network using the information stored in the numerical analysis database 240 and the measurement signal database 210.

  The model 230 is configured by extracting the numerical analysis result 6 necessary for calculating the model output 8 corresponding to the model input 7 from the numerical analysis database 240 using the model information data 5 and interpolating the result. The

  Further, in preparation for the difference between the characteristics of the physical model of the numerical analysis execution unit 220 and the characteristics of the plant 100, the control characteristics of the model 230 and the plant 100 are measured using the measurement signal 3 stored in the measurement signal database 210. Are configured so that the model 230 can be modified to match.

  The learning unit 260 learns how to generate the model input 7 so that the model output 8 simulated by the model 230 achieves the model output target value set in advance by the operator.

  At this time, the current model input is set as an input to the learning unit 260, and the model input change width is set as an output. Here, the input to the learning unit 260 is referred to as a state input, and the output is referred to as an operation amount change width.

  Learning information data 9 including constraint conditions used for learning and model output target values is stored in the learning information database 270.

  The learning unit 260 learns by using the state input as it is before starting the plant operation, and after starting the plant operation, uses pattern data stored in a pattern database 500 described later, which is a database provided in the control device 200. And patterning the state input.

  The learning data 10 that is the result of learning by the learning unit 260 is stored in a learning information database 270 that is a database provided in the control device 200.

  The learning data 10 includes information on the state input before the model input change, the operation amount change width in the state input, and the model output change width obtained as a result of the operation. The detailed function of the learning unit 260 will be described later.

  In the pattern generation unit 400 which is an arithmetic device provided in the control device 200, the learning unit 260 generates the pattern data 17 using the learning data 16 learned before the plant operation is started, and the pattern data 17 is transferred to the pattern database 500. save.

  After the operation of the plant 100 is started, the optimum pattern data 18 similar to the measurement signal 3 is selected from the pattern data stored in the pattern database 500, and the state input included in the measurement signal 3 is input by the learning unit 260. Learn by patterning.

  The learning data 11 obtained as a result of learning in the learning unit 260 and stored in the learning information database 270 is input to the learning result determination unit 300 that is an arithmetic device provided in the control device 200.

  In the learning result determination unit 300, when there are a plurality of pattern data 18 used for learning, learning data with the best control effect is selected from the learning data 11 of each pattern, and the pattern used is one Selects the learning data for that pattern.

  The learning data 12 selected by the learning result determination unit 300 is input to the operation signal generation unit 280.

  As shown in FIG. 1, the operator of the plant 100 includes an external input device 600 composed of a keyboard 601 and a mouse 602 and a data transmission / reception processing unit 612 that can transmit and receive data to and from the control device 200. By using the data processing device 610 and the image display device 620 provided, information stored in various databases provided in the control device 200 can be accessed.

  Further, the operator of the plant 100 uses the respective devices to set the setting parameters used in the numerical analysis execution unit 220, the learning unit 260, and the pattern generation unit 400, which are arithmetic devices provided in the control device 200. Can be entered.

  The data processing device 610 includes an external input interface 611, a data transmission / reception processing unit 612, and an external output interface 613.

  An input signal 61 of the data processing device generated by the external input device 600 is taken into the data processing device 610 via the external input interface 611. The data transmission / reception processing unit 612 of the data processing device 610 is configured to acquire the input / output data information 60 provided in the control device 200 according to the information of the data processing device input signal 62.

  Further, in the data transmission / reception processing unit 612, parameter setting used in the numerical analysis execution unit 220, the learning unit 260, and the pattern generation unit 400, which are arithmetic devices provided in the control device 200, according to the information of the data processing device input signal 62. Input / output data information 60 including values is output.

  The data transmission / reception processing unit 610 transmits a data processing device output signal 63 obtained as a result of processing the input / output data information 60 to the external output interface 613. Since the data processing device output signal 64 is sent from the external output interface 613 to the image display device 620 and displayed, information stored in various databases provided in the control device 200 is stored in the image display device 620. Can be displayed.

  In the plant control apparatus 200 according to the embodiment of the present invention described above, the measurement signal database 210, the numerical analysis database 240, the operation signal database 250, the learning information database 270, the control logic database 290, and the pattern database 500 are included. Although it arrange | positions so that the arithmetic unit of the control apparatus 200 may be comprised, all or some of these can also be arrange | positioned outside the control apparatus 200.

  Further, although the numerical analysis execution unit 220 is disposed inside the control device 200, it can also be disposed outside the control device 200.

  For example, the numerical analysis execution unit 220 and the numerical analysis database 240 described above may be arranged outside the control device 200, and the numerical analysis result 6 may be transmitted to the control device 200 via the Internet.

  Next, the learning unit 260 provided in the control device 200 will be described in detail with reference to FIG.

  As shown in FIG. 2, the learning unit 260 includes a model input generation unit 261, a preliminary learning unit 262, a pattern selection unit 263, a pattern conversion unit 264, and a re-learning unit 265.

  The model input generation unit 261 constituting the learning unit 260 applies the operation amount change width 19 to the current model input condition of the model 230, and the model input 7 and the state input 20 after the operation are this model input generation unit. 261.

  Further, the state input 20 is extracted by the model input generation means 261 from the measurement signal 3 of the plant 100 stored in the measurement signal database 210.

  The preliminary learning means 262 constituting the learning unit 260 uses the learning information data 9 stored in the learning information database 270 before the operation of the plant 100 is started, and the model output 8 simulated by the model 230 is the model output target. The method of generating the model input 7 is learned so as to achieve the value.

  At that time, the state input 20 is used as it is. The model input 7 is generated by applying the operation amount change width 19 learned by the preliminary learning unit 262 to the current model input condition in the model input generation unit 261, and inputs this to the model 230, and also after the operation. The state input 20 is input to the preliminary learning means 262.

  The learning data 10 obtained as a result of learning in the preliminary learning means 262 is stored in the learning information database 270.

  The pattern generation unit 400 generates pattern data based on the learning information 16 stored in the learning information database 270, and the pattern data 17 generated by the pattern generation unit 400 is stored in the pattern database 500.

  The pattern selection unit 263 constituting the learning unit 260 selects the pattern data 23 from the pattern data 18 stored in the pattern database 500 using the state input 20 as a clue, and inputs the pattern data 23 to the pattern conversion unit 264.

  The pattern conversion means 264 constituting the learning unit 260 patterns the state input 20 using the pattern data 23 selected by the pattern selection means 263, and reduces the input order.

  Further, the pattern conversion means 264 performs the reverse operation on the operation amount change width 21 learned by patterning to increase the input order.

  This patterning will be described with reference to FIG. 3. As shown in FIG. 3, in the pattern conversion means 264, the pattern data has an input information value of 1 used for patterning and an input information value of unused input of 0. Take the mode to do.

  Therefore, when patterning the state input 20, only the input whose pattern data information value is 1 is selected for the state input 20.

  Further, in the reverse operation, the operation amount change width of other inputs is derived by linearly interpolating between the inputs used in patterning with respect to the patterned operation amount change width 21 after learning.

  In FIG. 2, the re-learning means 265 constituting the learning unit 260 described above uses the learning information data 9 stored in the learning information database 270 after the operation of the plant 100 is started, and the model corrected by the measurement signal 3. The generation method of the model input 7 is learned so that the model output 8 simulated and operated at 230 achieves the model output target value.

  At that time, the pattern input means 264 patterns the state input 20. The patterned state input 22 is input to the re-learning unit 265 corresponding to the pattern, and as a result of learning by the re-learning unit 265, the patterned operation amount change width 21 is output from the re-learning unit 265. Is done.

  This is converted again to the manipulated variable change width 19 by the pattern converting means 264 and input to the model input generating means 261 to generate the model input 7 and the state input 20 after the operation.

  With the configuration described above, the learning unit 260 implements preliminary learning for pattern generation before the plant 100 starts operation and relearning by patterning state input after the plant 100 starts operation.

  FIG. 4 is a flowchart showing the operation of the control device 200 that controls the plant according to one embodiment of the present invention described with reference to FIGS. 1 to 3.

  In the flowchart of FIG. 4, steps 1000, 1010, 1020, 1030, 1040, and 1050 are executed in combination. Hereinafter, each step will be described.

  After the operation of the control device 200 that controls the plant 100 is started, first, in step 1000, numerical analysis is performed using the numerical analysis execution unit 220 of the control device 200, and the numerical analysis data 4 is transmitted to and stored in the numerical analysis database 240. To do.

  Next, in Step 1010, preliminary learning of the plant operation method is executed using the learning unit 260 and the model 230 of the control device 200, and the learning data 10 is transmitted and stored in the learning information database 270.

  Next, in Step 1020, the pattern generation unit 400 of the control device 200 is used to generate the pattern data 17 using the learning data 16 by preliminary learning stored in the learning information database 270, and the pattern data 17 is transmitted to the pattern database 500. save. The above operation is executed before the operation of the plant 100 is started.

  After the operation of the plant 100 is started, in step 1030, the measurement signal 1 of the plant 100 is input to the control device 200 using the external input interface 201, and is transmitted and stored in the measurement signal database 210.

  In step 1040, the state input extracted from the acquired measurement signal 3 is patterned using the pattern data 18 stored in the pattern database 500, and the learning unit 260 performs relearning.

  After learning in the learning unit 260, the learning result determination unit 300 of the control device 200 is used to select appropriate learning data 12 from the learning data 11 based on the used pattern, and from the learning result determination unit 300 to the operation signal generation unit 280. Output to.

  In the next step 1050, the operation signal generator 280 is used to generate the operation signal 14 using the learning data 12 and the control logic data 13, and the operation signal is generated from the operation signal generator 280 via the external input interface 202. 15 is output to the plant 100.

  The plant 100 is controlled by repeatedly executing the operations in steps 1030 to 1050 described above each time a measurement signal is input.

  Next, detailed operations in steps 1010, 1020, and 1040 in FIG. 4 will be described in detail with reference to the flowchart.

  FIG. 5 is a flowchart showing the preliminary learning operation in step 1010.

  As shown in FIG. 5, the flowchart of the preliminary learning operation is executed by combining Steps 2000, 2010, 2020, 2030, 2040, 2050, 2060, 2070, and 2080.

  Each step will be described in detail below.

  In step 2000, an initialization count I, which is a value indicating the number of repetitions of steps 2010 to 2060, is initialized (set to I = 1).

  Next, in step 2010, an initial value for state input is set. Any model input value can be selected as the initial value of the state input.

  In step 2020, the number of operations J, which is a value indicating the number of repetitions of steps 2030 to 2050, is initialized (set to J = 1).

  Next, in Step 2030, the amount of change in the manipulated variable of the model input is derived using the preliminary learning means 262 of the learning unit 260 constituting the control device 200.

  In step 2040, the model input is generated by using the model input generation means 261 of the learning unit 260, the operation amount variation is added to the state input, and the model 230 is operated to update the state input.

  In step 2050, using the model output obtained as a result of the operation of the model 230 as a clue, the operation method of the model 230 is learned by using a neural network and various learning algorithms represented by reinforcement learning.

  Step 2060 is a branch, and the number of operations J is compared with a predetermined threshold. If J is smaller than the threshold, J is incremented by 1 and the process returns to Step 2030. Conversely, if J is larger than the threshold, Advances to step 2070 which is a branch.

  In step 2070, the initialization count I is compared with a predetermined threshold value. If I is smaller than the threshold value, I is incremented by 1, and then the process returns to step 2010. Proceed to 2080.

  In step 2080, the result learned by the learning unit 260 is transmitted to and stored in the learning information database 270, and the process proceeds to a step of terminating the preliminary learning operation.

  With the above operation, in the preliminary learning, learning is started with an arbitrary model input as an initial value of the state input in step 2010, so that an operation method for reaching the model output target value from an arbitrary model input condition can be obtained.

  FIG. 6 is a flowchart showing the pattern generation operation in step 1020.

  As shown in FIG. 6, the flowchart of the pattern generation operation is executed by combining Steps 4000, 4010, 4020, and 4030.

  Each step will be described in detail below with reference to FIGS.

  In step 4000, a data reference number I which is a value indicating the number of repetitions of steps 4010 to 4020 is initialized (I = 1 is set).

  Next, in step 4010, a pattern is generated while referring to learning data by preliminary learning stored in the learning information database 270 according to I. The detailed operation of the pattern generation algorithm will be described later.

  FIG. 7 shows an aspect of learning data stored in the learning information database 270 of the control device 200.

  As shown in FIG. 7, the learning information database 270 stores a state input value, an operation amount change width value in the state input, and a model output change width value obtained as a result of the operation.

  The pattern is generated based on the operation amount change width in the learning data. In FIG. 7, S_0001 is a number assigned to distinguish state inputs.

  In step 4020 in FIG. 6, the generated pattern and the state input value of the referenced learning data are transmitted and stored in the pattern database 500.

  FIG. 8 shows an aspect of learning data stored in the pattern database 500 of the control device 200.

  As shown in FIG. 8, the pattern database 500 stores pattern data and state input values referred to when the pattern is generated.

  When the generated pattern is saved, if it matches the already saved pattern data, it is saved as a state input corresponding to the pattern data.

  Therefore, a case where a plurality of state inputs are associated with one pattern data occurs. In FIG. 8, P_0001 is a number assigned to distinguish pattern data.

  Step 4030 in FIG. 6 is a branch, and the data reference number I is compared with a predetermined threshold value. If J is smaller than the threshold value, J is incremented by 1 and the process returns to step 4010. If it is smaller, the process proceeds to the step of ending the pattern generation operation.

  Next, the detailed operation of the pattern generation algorithm in step 4010 will be described with reference to the flowchart of FIG. 9 and FIG.

  In the pattern generation algorithm, in order to keep the input order to be patterned to the minimum necessary value, the search starts from the pattern input order 1 as shown in FIG. 10, and the input order is increased until the end condition is satisfied. The pattern is generated by repeating the search.

  In order to make the pattern search more efficient, an evolutionary search algorithm is used for the search. That is, an optimal pattern is searched by using a database expression of a pattern as a gene (solution candidate) and performing a genetic operation such as crossover or mutation on a plurality of generated solution candidates.

  As shown in FIG. 9, the flowchart of the operation of Step 4010 is executed by combining Steps 4011, 4012, 4013, 4014, 4015, 4016, 4017, 4018, and 4019.

  Each step will be described in detail below.

  In step 4011, J, which is a value indicating the input order used for patterning, is set to initialization (J = 1).

  In step 4012, the generation number K, which is a value indicating the number of repetitions of steps 4014 to 4015, is initialized (set to K = 1).

  Next, in step 4013, L solution candidates are generated under the constraint condition that the input order at which the pattern data information value is 1 is J.

  In step 4014, a patterning error is calculated for each solution candidate.

  As shown in FIG. 10, the patterning error is obtained as an error between the operation amount change width of the learning data to be referred to and the operation amount change width obtained by linear interpolation of the input used for patterning.

  In step 4015, based on the calculated patterning error, a genetic operation such as crossover or mutation is performed on the solution candidate group under the constraint conditions to generate a new solution candidate.

  Step 4016 is a branch. If the number of generations K is less than or equal to a predetermined threshold, K is incremented by 1 and then the process returns to step 4014. In the opposite case, the process proceeds to step 4017.

  In step 4017, a solution candidate that minimizes the patterning error is selected from the solution candidate group that has passed the predetermined number of generations, and the process proceeds to step 4018.

  Step 4018 is a branch. If the patterning error of the solution candidate selected in Step 4017 is greater than or equal to a predetermined threshold value, or if the input order J is less than or equal to a predetermined threshold value, J is set. After 1 is added, the process returns to Step 4012. In the opposite case, the process proceeds to Step 4019.

  In step 4019, the selected solution candidate is used as a generation pattern, which is transmitted to and saved in the pattern database 500, and the process proceeds to the step of terminating the operation of the pattern generation algorithm.

  Through the above operation, the generated pattern data is stored in the pattern database 500.

  An operator of the plant 100 can display the information on the image display device 620 using the data processing device 610.

  As a result, the operator can know what state input patterning is currently possible. In addition, threshold values for setting parameters J and K can be input through the screen display device 620.

  FIG. 11 is a flowchart showing the relearning operation in step 1040.

  As shown in FIG. 11, the flowchart of the relearning operation is executed by combining steps 2100, 2110, 2120, 2130, 2140, 2150, 2160, 2170, and 2180.

  Each step will be described in detail below.

  In step 2100, the state input input to the re-learning unit 265 is extracted from the measurement signal using the model input generation unit 261 of the learning unit 260 constituting the control device 200.

  Next, in step 2110, the minimum state input error between the extracted state input and the pattern data stored in the pattern database 500 of the control device 200 is calculated using the pattern selection unit 263 of the learning unit 260.

  As shown in FIG. 8, the pattern database 500 stores a plurality of state input values for each pattern, and the minimum state input error is obtained as the minimum value of the error between these and the current state input value.

  Since the minimum state input error is the similarity between the current state input and the state input experienced in the preliminary learning, a pattern corresponding to the state input having a high similarity with the current state input value can be selected using this as a clue.

  As a result, by using the selected pattern, it is possible to learn an operation amount change width that can obtain a control effect equivalent to the preliminary learning result for the current state input.

  Step 2120 is a branch, and it is determined whether or not there is a value that is equal to or smaller than a preset threshold with respect to the minimum state input error of each pattern calculated in Step 2110. If there is, the process proceeds to Step 2130. If not, go to Step 2140.

  In step 2130, all patterns whose minimum state input error is equal to or smaller than the threshold value are selected, and the process proceeds to step 2150.

  In step 2140, only the pattern having the smallest minimum state input error among all patterns is selected, and the process proceeds to step 2150.

  Next, in step 2150, I, which is a value indicating the number of repetitions of step 2160, is initialized (set to I = 1).

  In step 2160, the model operation method of the model 230 is learned using the relearning means 265 of the learning unit 260 corresponding to the selected pattern.

  Step 2170 is a branch. If I is equal to or less than the number of selected patterns, I is incremented by 1, and the process returns to step 2160 to learn using the re-learning means 265 corresponding to another selected pattern. In the opposite case, the process proceeds to Step 2180.

  In step 2180, the learning result determination unit 300 of the control device 200 is used to determine an appropriate operation method from the re-learning result based on the selected pattern, which is output to the operation signal generation unit 280 of the control device 200 for re-learning. Proceed to the step of terminating the operation.

  With the above operation, in re-learning, an operation method equivalent to the preliminary learning result can be learned in a short time by using a pattern selected based on the preliminary learning result for the current state input.

  In addition, a threshold value of a minimum state input error that is a setting parameter can be input through the screen display device 620.

  Next, the detailed operation of the relearning algorithm in step 2160 will be described in detail with reference to the flowchart of FIG.

  As shown in FIG. 12, the flowchart of the operation of the re-learning algorithm is executed by combining steps 2200, 2210, 2220, 2230, 2240, 2250, 2260, 2270, 2280, 2290, and 2300.

  Each step will be described in detail below.

  In step 2200, an initialization count J, which is a value indicating the number of repetitions of steps 2210 to 2280, is initialized (set to J = 1).

  Next, in step 2210, the state input is initialized to an input value extracted from the measurement signal 3.

  In step 2220, the number of operations K, which is a value indicating the number of repetitions of steps 2230 to 2270, is initialized (set to K = 1).

  Next, in Step 2230, the re-learning means 265 of the learning unit 260 corresponding to the pattern to be used is used to derive the operation amount change width of the modeled model input.

  In step 2240, the pattern conversion means 264 of the learning unit 260 is used to linearly interpolate the manipulated variable variation width of the modeled model input to derive the manipulated variable variation width.

  In step 2250, the model input is generated by using the model input generation means 261 of the learning unit 260, the model is operated by adding the operation amount change width to the state input, and the state input is updated.

  In step 2260, the pattern input unit 264 of the learning unit 260 is used to pattern the state input according to the pattern data to be used.

  In step 2270, a model operation method is learned using a neural network and various learning algorithms represented by reinforcement learning, using the model output obtained as a result of the model operation as a clue.

  Step 2280 is a branch, and the number of operations K is compared with a predetermined threshold value. If K is smaller than the threshold value, 1 is added to K and then the process returns to Step 2230. Advances to step 2290 which is a branch.

  In step 2290, the initialization count J is compared with a predetermined threshold value. If J is smaller than the threshold value, J is incremented by 1, and then the process returns to step 2210. Proceed to 2300.

  In step 2300, the learning result is transmitted and stored in the learning information database 270 of the control device 200, and the process proceeds to the step of ending the relearning operation.

  With the above operation, in the re-learning, the operation method when the current state input is set as the initial value is learned using the patterned state input.

  Above, description of the detailed operation | movement of step 1010, 1020, and 1040 shown in FIG. 4 is complete | finished.

  Next, FIG. 13 to FIG. 13 show screens displayed on the image display device 620 for displaying the data processing device output signal 64 output from the data processing device 610 constituting the plant control device 200 according to the embodiment of the present invention. 21 will be described.

  FIGS. 13-20 is an example of the screen displayed on the image display apparatus 620 attached to the plant control apparatus 200 shown in FIG.

  FIG. 13 shows an example of the initial screen. When the screen of FIG. 13 is displayed on the image display device 620, the mouse 602 is operated to place the cursor on the button, and the mouse 602 is clicked to click the button. Can be selected (pressed).

  When the “numerical analysis” button 701, “database reference” button 702, “patterning / learning execution” button 703, and “operation execution” button 704 shown in FIG. 13 are selected, respectively, FIG. 14 and FIG. 16 and 17 are displayed on the screen display device 620. When the “end” button 705 is selected, the initial screen is ended.

  In FIG. 14, by selecting an “analysis condition setting” button 711, various analysis conditions necessary for executing a calculation by the numerical analysis execution unit 220 of the control device 200 can be input / set.

  When the “numerical analysis execution” button 712 is selected, the numerical analysis execution unit 220 can start the calculation. When the “return” button 713 is selected, the process returns to FIG.

  In FIG. 15, it is possible to select in which database the information stored in the image display device 620 is displayed.

  “Measurement signal database” button 721, “Operation signal database” button 722, “Numeric analysis database” button 723, “Learning information database” button 724, “Control logic database” button 725, “Pattern database” By selecting the button 726, the measurement signal database 210, the operation signal database 250, the numerical analysis database 240, the learning information database 270, the control logic database 290, and the pattern database 500 of the control device 200 can be accessed.

  Information on each database can be displayed on the image display device 620, and information on the database can be added / changed / erased. When the “return” button 727 is selected, the process returns to FIG.

  In FIG. 16, by selecting the “preliminary learning” button 731, the preliminary learning screen illustrated in FIG. 18 is displayed on the screen display device 620.

  Also, by selecting the “relearn” button 732 in FIG. 16, the relearning screen shown in FIG. 19 is displayed on the image display device 620. When the “return” button 733 is selected, the process returns to FIG.

  In FIG. 17, the model input 741 before the operation, the model output 742, the model input 743 after the operation, the model output 744, and the learned operation amount change width 745 are displayed as guidance, and the operator of the plant 100 responds to the guidance display. The user can select whether to perform the operation.

  If the operation is to be executed, the “Yes” button 746 is selected. If the operation is not to be executed, the “No” button 747 is selected.

  In the display screen shown in FIG. 17, since the plant can be operated after the relationship between the pattern and the operation signal is confirmed on the screen, the reliability of the plant operation can be improved.

  That is, the plant control method shown in FIG. 17 will be described in detail. In the plant control method of a control device that controls a plant using a model simulating a plant, the control device is configured as shown in FIG. It has a pattern database that stores a state input pattern obtained by patterning a plurality of operation signals input to the model, and stores a plurality of operation signals input to the model and outputs from the model as shown in FIG. As shown as a model input 743 and a model output 744 after operation, a plurality of operation signals stored in the learning information database and outputs from the model are output to a display device, and learning is performed. As shown as the manipulated variable change width 745, the operation signal to the plant based on the plurality of operation signals input to the model and the parameter The state input pattern stored in the database is overlaid on the display device and output to the display device. As shown as the button 746 to be selected when the operation is executed, the plant is controlled based on the permission to execute the operation on the plant. It is comprised so that a plant may be controlled by the operation signal to.

  In FIG. 18 described above, selecting the “learning start” button 751 causes the preliminary learning means 262 of the learning unit 260 and the model 230 of the control device 200 to operate according to the flowchart shown in FIG. Can be executed.

  Further, by selecting the “pattern generation” button 752, the pattern generation screen shown in FIG. 20 is displayed on the screen display device 620.

  When the “return” button 753 is selected, the process returns to FIG.

  In FIG. 19, the setting value of the minimum state input error threshold used in the flowchart shown in FIG. 11 can be input to the data input field 761 as the setting value at the time of relearning.

  When the “execute” button 762 is selected after the set value is input in the data input field 761, the learning unit 260 and the model 230 of the control device 200 are operated according to the flowchart shown in FIG. it can.

  When the “return” button 763 is selected, the process returns to FIG.

  In FIG. 20, as the setting value at the time of pattern generation, the setting value of the patterning error threshold used in the flowchart shown in FIG. 9 is set in the data input column 771, and the setting value of the pattern input order threshold is set in the data input column 772. Can be entered.

  When the “execute” button 773 is selected after the set values are entered in the data input fields 771 and 772, the pattern generation unit 400 of the control device 200 is operated according to the flowchart shown in FIG. 6 to execute pattern generation. it can.

  When the “return” button 774 is selected, the process returns to FIG.

  Above, description about the screen displayed on the image display apparatus 620 is complete | finished.

  As is apparent from the above description, in the embodiment of the present invention, as shown in detail using the flowcharts in FIGS. 11 and 12, the state input input to the learning unit of the plant control device is a model. Learning can be speeded up by patterning with fewer inputs than the input order and reducing the model search space.

  Further, as shown in detail using the flowchart in FIG. 5, a pattern is generated using a preliminary learning result that has been learned without reducing the input order in advance before the plant operation is started. As shown in Fig. 4, after the plant operation starts, the similarity between the current state input and the state input included in the preliminary learning result is compared, and the operation method is learned by selecting the pattern generated from the state input with a high degree of similarity. By doing so, a desired learning result can always be obtained.

  Further, as shown in detail using the flowchart in FIG. 9, the pattern input order and pattern information are automatically generated by an evolutionary search algorithm, so that erroneous patterning can be performed without human intervention. It is possible to generate a pattern to avoid.

  Further, the function for referring to the information stored in each database by the image display device, and the setting parameters used by the learning unit and the pattern generation unit are shown through the image display device as shown in detail in FIGS. By providing the input function, the plant operator can visually confirm the patterning and the effect of the patterned learning.

  Next, the case where the plant control apparatus 200 which is one embodiment of the present invention shown in FIGS. 1 to 20 is applied to a thermal power plant 100 equipped with a boiler will be described. Needless to say, the plant control device 200 of the above-described embodiment of the present invention can also be used when controlling a plant other than the thermal power plant.

  FIG. 21 is a diagram illustrating a schematic configuration of a thermal power plant including a boiler. First, the mechanism of a plant in a thermal power plant equipped with a boiler will be described.

  In the boiler 101 constituting the thermal power plant shown in FIG. 21, coal serving as fuel is pulverized by a mill 110 to form pulverized coal in the boiler 101 together with primary air for transporting coal and secondary air for combustion adjustment. The boiler 101 is put into the boiler 101 through the installed burner 102, and fuel coal is burned inside the furnace of the boiler 101.

  The fuel coal and the primary air are led from the pipe 134 and the secondary air is led from the pipe 141 to the burner 102.

  Further, after-air for two-stage combustion is introduced into the boiler 101 through an after air port 103 installed in the boiler 101. This after air is guided from the pipe 142 to the after air port 103.

  High-temperature combustion gas generated by burning fuel coal inside the furnace of the boiler 101 flows downstream along the path indicated by the arrow in the furnace of the boiler 101, and the heat exchanger 106 disposed in the boiler 101. After passing through and exchanging heat, it becomes combustion exhaust gas and is discharged from the boiler 101 and flows down to the air heater 104 installed outside the boiler 101.

  The combustion exhaust gas that has passed through the air heater 104 is then released from the chimney to the atmosphere after removing harmful substances contained in the combustion exhaust gas with an exhaust gas treatment device (not shown).

  The feed water circulating in the boiler 101 is guided to the boiler 101 from a condenser (not shown) installed in the turbine 108 via the feed water pump 105, and the furnace 101 of the boiler 101 is installed in the heat exchanger 106 installed in the furnace of the boiler 101. It is heated by the combustion gas flowing down the inside and becomes high-temperature and high-pressure steam.

  In this embodiment, the number of heat exchangers 106 is shown as one, but a plurality of heat exchangers may be arranged.

  The high-temperature and high-pressure steam generated in the heat exchanger 106 is guided to the steam turbine 108 via the turbine governor valve 107, and the steam turbine 108 is driven by the energy of the steam, and the generator 109 connected to the steam turbine 108 is connected to the generator 109. Rotate to generate electricity.

  Next, the primary air and the secondary air introduced into the furnace 101 of the boiler 101 from the burner 102 installed in the furnace of the boiler 101, and the after air port 103 installed in the furnace of the boiler 101 are introduced into the furnace of the boiler 101. The after-air route will be described.

  The primary air is guided from the fan 120 to the pipe 130, and is branched into a pipe 132 that passes through the inside of the air heater 104 and a pipe 131 that bypasses the air heater 104, and flows down the pipe 132 and the pipe 131. The primary air joins again in the pipe 133 and is guided to the mill 110.

  Air passing through the air heater 104 is heated by the combustion exhaust gas discharged from the furnace of the boiler 101.

  The primary air is used to convey coal (pulverized coal) generated in the mill 110 to the burner 102 through the pipe 133.

  The secondary air and the after air are led from the fan 121 to the pipe 140 and flow down the pipe 140 passing through the inside of the air heater 104 and heated, and then the secondary air pipe 141 on the downstream side of the pipe 140, The pipe is branched to an after-air pipe 142 and led to a burner 102 and an after-air port 103 installed in the furnace of the boiler 101, respectively.

  The control apparatus 200 of the thermal power plant 100 provided with the boiler according to the present embodiment reduces the NOx and CO concentrations in the exhaust gas of the boiler, and the amount of air to be input from the burner 102 to the boiler 101 and the boiler from the after air port 103. 101 has a function of adjusting the amount of air to be supplied to the apparatus 101.

The thermal power plant 100 is provided with various measuring devices that detect the operating state of the thermal power plant 100, and the plant measurement signals acquired from these measuring devices are sent to the control device 200 as measurement signals 1. As various measuring instruments for detecting the operating state of the thermal power plant 100 to be transmitted, for example, in FIG. 2, a flow measuring instrument 150, a temperature measuring instrument 151, a pressure measuring instrument 152, a power generation output measuring instrument 153, and an O ₂ concentration and A concentration measuring device 154 for measuring CO concentration is shown.

  The flow rate measuring device 150 measures the flow rate of the feed water supplied from the feed water pump 105 to the boiler 101. Further, the temperature measuring device 151 and the pressure measuring device 152 are used to supply steam generated by heat exchange with the combustion gas flowing down the boiler 101 in the heat exchanger 106 disposed in the boiler 101 to the steam turbine 108. Measure temperature and pressure respectively.

  The amount of power generated by the generator 109 rotated by the steam turbine 108 driven by the steam generated in the heat exchanger 106 is measured by a power generation output measuring device 153.

In addition, information on the concentration of components (CO, NOx, etc.) contained in the combustion gas flowing down the boiler 101 is the O ₂ concentration and / or the CO concentration provided in the boiler outlet channel on the downstream side of the boiler 101. It is measured by a concentration measuring device 154 that measures

  In general, many measuring instruments other than those shown in FIG. 2 are arranged in the thermal power plant 100, but the illustration is omitted here.

  FIG. 22 is a partially enlarged view showing an air heater 104 installed on the downstream side of the boiler 101 constituting the thermal power plant 100 and piping arranged in the air heater 104.

  As shown in FIG. 22, a secondary air pipe 141 and an after-air pipe 142 branched on the downstream side of the pipe 140 arranged inside the air heater 104, and a pipe arranged inside the air heater 104. 132, and air dampers 162, 163, 161, and 160 are disposed in the pipe 131 bypassing the air heater 104, respectively.

  By operating these air dampers 160 to 163, the area through which air passes in the pipes 131, 132, 141, 142 is changed, and the flow rate of air passing through these pipes 131, 132, 141, 142 is individually adjusted. To do.

  Then, using the operation signal 15 generated by the control device 200 that controls the thermal power plant 100 and output to the thermal power plant 100, the feed water pump 105, the mill 110, the air dampers 160, 161, 162, 163, etc. Operate the equipment.

  In addition, in the control apparatus 200 of the thermal power plant 100 which is a present Example, the equipment which adjusts the state quantity of thermal power plants, such as the water supply pump 105, the mill 110, the air dampers 160, 161, 162, 163, is an operation end. A command signal required to operate the signal is called an operation signal.

  FIG. 23 is a flowchart showing the procedure of the thermal power plant control method in the control device 200 of the thermal power plant 100 including the boiler according to the embodiment of the present invention shown in FIG.

  23, in the control apparatus 200 of the thermal power plant 100 provided with the boiler according to the present embodiment, the control of the thermal power plant 100 is executed by combining Steps 2181 and 2182.

  In step 2181, the learning output by each pattern selected by the pattern selection means 263 constituting the learning unit 260 of the control device 200 that controls the thermal power plant 100 is referred to, and the model output change after operation with respect to the current state input The total of the improved values of CO and NOx, which are widths, is compared, and the pattern that maximizes the total improved value is selected.

  If the number of patterns selected by the pattern selection unit 263 is 1, the process proceeds to step 2182 without executing the above process.

  In step 2182, the learning data 12 of the pattern selected by the pattern selection unit 263 is transmitted to the operation signal generation unit 280 constituting the control device 200, and the operation of the learning result determination unit 300 constituting the control device 200 is terminated. Proceed to step.

  As a result of the above-described operation, the control device 200 can generate an operation signal using learning data that provides the best improvement in CO and NOx as a result of relearning.

  FIG. 24 shows an example of a screen displayed on the image display device 620 in the control device 200 of the thermal power plant 100 provided with the boiler according to the embodiment of the present invention. The screen shown in FIG. This is a screen corresponding to the example.

  In the example of the screen shown in FIG. 24, the air amount of the burner 102 of the boiler 101 used as the model input 7 input to the model 230 constituting the control device 200, the air amount of the air port 103, the control target of the model 230 The burner 102 and the airport learned by the learning unit 265 by the guidance display before and after the operation of the CO, NOx concentration, and the pattern conversion means 264, which are used as the model output 8 output by simulating the thermal power plant. The model input 7 displays the distribution of the air amount of the burner 102 / air port 103 installed before and after the boiler 101 in association with the boiler structure diagram.

  In FIG. 24, the burner 102 and the air port 103 are provided in one stage both before and after the boiler 101, and five air inlets are provided in each stage. However, these numbers may be arbitrarily changed. You can also.

  The operator of the plant can check the patterned operation amount change width and the control effect by the learned operation method while observing the screen shown in FIG. 24, and can determine the presence or absence of the operation.

  Further, as schematically shown in FIG. 25, as an example of patterning in the control device 200 of the thermal power plant 100 including the boiler according to the embodiment of the present invention, 1) for each burner 102 / air port 103 of the boiler 101 An example of patterning using either one of grouping operation ends, 2) grouping operation ends before / after boiler 101, and 3) grouping all operation ends of boiler 101 is conceivable.

  As is clear from the above description, the operator of the thermal power plant 100 equipped with the boiler takes into account the type of operation end of the boiler provided in the thermal power plant to be controlled, the number of operation ends, combustion characteristics, and the like. The thermal power plant can be controlled by selecting an appropriate patterning means.

  According to the embodiment of the present invention, a model is corrected based on a measurement signal obtained by measuring the operating state of the plant, and learning that is learned again using the corrected model is executed at high speed to perform a plant control algorithm. And a plant control device that controls the plant with high accuracy can be realized.

  The present invention is applicable to a plant control device, particularly a thermal power plant control device.

The control block diagram which shows the whole structure of the control apparatus of the plant which is one Example of this invention. The control block diagram which shows the learning part in the control apparatus of the plant which is one Example of this invention shown in FIG. Explanatory drawing explaining the operation | movement of patterning in the control apparatus of the plant which is one Example of this invention shown in FIG. The flowchart which shows operation | movement of the control apparatus of the plant which is one Example of this invention shown in FIG. The flowchart which shows the operation | movement of the preliminary learning in the flowchart which showed operation | movement of the control apparatus of the plant which is one Example of this invention shown in FIG. The flowchart which shows the operation | movement of the pattern generation in the flowchart which showed operation | movement of the control apparatus of the plant which is one Example of this invention shown in FIG. Explanatory drawing which shows the one aspect | mode of the information preserve | saved in the learning information database in the control apparatus of the plant which is one Example of this invention shown in FIG. Explanatory drawing which shows the one aspect | mode of the information preserve | saved in the pattern database in the control apparatus of the plant which is one Example of this invention shown in FIG. 5 is a flowchart illustrating in detail a pattern search algorithm which is a part of pattern generation operation in the flowchart illustrating the operation of the plant control apparatus according to the embodiment of the present invention illustrated in FIG. 4. Explanatory drawing explaining the pattern search algorithm which calculates the patterning error in the flowchart which showed the operation | movement of the control apparatus of the plant which is one Example of this invention shown in FIG. The flowchart which shows the operation | movement of the relearning in the flowchart which showed operation | movement of the control apparatus of the plant which is one Example of this invention shown in FIG. 5 is a flowchart illustrating in detail a part of the relearning operation in the flowchart illustrating the operation of the plant control apparatus according to the embodiment of the present invention illustrated in FIG. 4. The example which showed the example of the initial screen displayed on the image display apparatus attached to the control apparatus of the plant by one Example of this invention shown in FIG. The example which showed the screen displayed on an image display apparatus, when performing numerical analysis in the control apparatus of the plant by one Example of this invention shown in FIG. The example which showed the screen displayed on an image display apparatus, when referring to a database in the control apparatus of the plant by one Example of this invention shown in FIG. An example of the screen displayed on an image display apparatus, when patterning and learning are performed in the control apparatus of the plant by one Example of this invention shown in FIG. An example of the screen displayed on an image display apparatus when performing operation in the control apparatus of the plant by one Example of this invention shown in FIG. FIG. 2 is an example of a screen displayed on the image display device when performing preliminary learning and pattern generation in the plant control apparatus according to the embodiment of the present invention shown in FIG. 1. FIG. An example of the screen displayed on an image display apparatus, when re-learning is performed in the control apparatus of the plant by one Example of this invention shown in FIG. An example of the screen displayed on an image display apparatus, when performing pattern generation in the control apparatus of the plant by one Example of this invention shown in FIG. The schematic block diagram which shows the thermal power plant provided with the boiler of the control object to which the control apparatus of the plant of the Example of this invention shown in FIG. 1 is applied. The enlarged view of the air heater part in the boiler with which the thermal power plant of the control object shown in FIG. 21 was equipped. The flowchart which shows a part of the content of the relearning in the control apparatus of the thermal power plant provided with the boiler which is one Example of this invention shown in FIG. FIG. 22 is an example of a screen displayed on the image display device when an operation is performed in the control device for the thermal power plant including the boiler according to the embodiment of the present invention illustrated in FIG. 21. Explanatory drawing which shows the example of the patterned grouping in the control apparatus of the thermal power plant provided with the boiler which is one Example of this invention shown in FIG.

Explanation of symbols

  100: Plant, 200: Controller, 201: External input interface, 202: External output interface, 210: Measurement signal database, 220: Numerical analysis execution unit, 230: Model, 240: Numerical analysis database, 250: Operation signal database, 260: learning unit, 270: learning information database, 280: operation signal generation unit, 290: control logic database, 300: learning result determination unit, 400: pattern generation unit, 500: pattern database, 600: external input device, 601: Keyboard: 602: Mouse, 610: Data processing device, 611: External input interface, 612: Data transmission / reception processing unit, 613: External output interface, 620: Image display device.

Claims (10)

In a plant control device including an operation signal generation unit that calculates an operation signal that serves as a control command for a plant to be controlled using a measurement signal obtained by measuring the operation state of the plant, the control device includes: A measurement signal database in which measurement signals are stored, an operation signal database in which operation signals for the plant are stored, a numerical analysis execution unit that analyzes the operation characteristics of the plant, and a numerical analysis from the numerical analysis execution unit A model that simulates the plant control characteristics when an operation signal is given to the plant based on the information of the result , and the model output that achieves the model output target value by the model output that is simulated by the model using the model when learning how to produce, before the plant operation start learning using as a state input, which is input to the learning section, plant Pattern after start rolling a learning unit that learns by patterning the state input which is input to the learning unit using the pattern data stored in the pattern database, the learning unit stored in the pattern database It is configured to select pattern data from the data, pattern the state input using the selected pattern data and reduce the input order, and learning information data obtained as a result of learning in the learning unit is stored Learning information database, a control logic database in which information used for the operation signal output from the operation signal generation unit is stored , learning data learned before the plant operation is started in the learning unit, and the learning information a pattern generator for generating a pattern data based on the learning data stored in the database, A pattern database in which patterns data generated by the serial pattern generation unit is stored, the pattern data used for learning in the case of multiple effects to accurately control the plant from the training data of each pattern is good training data A control apparatus for a plant, comprising: a learning result determination unit that selects the operation signal, and is configured to calculate the operation signal from the operation signal generation unit based on learning data selected by the learning result determination unit . The plant control apparatus according to claim 1, wherein the learning unit generates an input to be applied to the model and extracts a model input from a measurement signal of the plant, and learning information data before starting plant operation. A preliminary learning means for learning a model input generation method so that the model output simulated by the model using the learning information data stored in the base achieves the target value; similar learning means has stored the generated pattern using a pattern generation unit relative to learning information data learned in the preliminary learning, from the pattern data stored in the pattern database after the plant operation starts measurement signal a pattern selecting means for selecting pattern data, pattern selected by the pattern selecting means Is patterned to reduce the input order information value using Ndeta selects the only condition input to be 1, linear interpolation between input using the pattern with respect to patterned MV change width after learning a pattern conversion means you increase the input order to derive the manipulated variable change the width of the other input and were simulated calculation model using learning information data after plant operation start is stored in the learning information database Re-learning means for patterning the state input by the pattern conversion means and learning by inputting the patterned state input so that the model output achieves its target value. A plant control device.   The plant control device according to claim 1, wherein the pattern generation unit configuring the control device includes the model operation amount change width information included in the learning information data learned using the preliminary learning unit before the plant operation is started. On the other hand, in order to generate pattern data, the search is started from the input order 1 at first, and the search is repeatedly performed while increasing the input order by 1 until the end condition is satisfied, and a plurality of patterns as search means for the pattern data A plant control device having a function of searching for pattern data by generating data candidates, performing genetic operations such as crossover and mutation on the data candidates, and overlapping the number of searches.   The plant control apparatus according to claim 2, wherein the learning unit configuring the control apparatus calculates a similarity between the current state input and the state input included in the preliminary learning result in the pattern selection unit after the plant operation is started. A plant control apparatus characterized by adding a learning method of operation by selecting a pattern generated from a state input having a high degree of similarity.   The plant control apparatus according to claim 1, wherein the control apparatus is stored in a measurement signal database, an operation signal database, a control logic database, a learning information database, a numerical analysis database, and a pattern database, respectively. A plant control apparatus comprising a data processing device and an image display device for displaying information on the screen.   The plant control device according to claim 5, wherein setting parameters used in the learning unit and pattern generation unit constituting the control device are input via an input unit connected to the data processing device. A plant control apparatus characterized by comprising.   2. The plant control apparatus according to claim 1, wherein the plant to be controlled is a thermal power plant including a boiler, and an operation end to which the operation signal for the plant is output is generated by a combustion reaction between fossil fuel and air. A burner that generates a high-temperature gas, and an air port that supplies air to the combustion gas on the downstream side in the flow direction of the combustion gas generated by mixing the coal fuel and air supplied from the burner. The measurement signal is a measurement signal of the carbon monoxide concentration and the nitrogen oxide concentration obtained by measuring the carbon monoxide concentration and the nitrogen oxide concentration of the combustion gas burned in the boiler, and the learning result constituting the control device. The determination unit includes a concentration of carbon monoxide and nitrogen oxides after the operation from the learning information data according to the selected pattern. Every operation total plant control system, characterized in that it is configured to generate the operation signal to select the one that the maximum improvement value indicating the difference before and after.   8. The plant control apparatus according to claim 7, wherein the state input parameters to be patterned by the learning result determination unit constituting the control apparatus are individually set from the burner and the air port installed in the boiler. A plant control device characterized by the flow rate of air to be introduced into the plant.   In the plant control apparatus according to claim 7 or 8, in the learning result determination unit constituting the control apparatus, grouping of state inputs to be patterned is performed for each burner / airport installed in the boiler. A plant control apparatus characterized by using any one of grouping, grouping before / after boiler cans, and grouping the boilers as a whole. In the plant control system according to claim 1, wherein the plurality of operation signals stored in the learning information database and displays the output from the model, to the plant based on the plurality of operation signal input to the model A plant control apparatus comprising: a display device that outputs and displays the operation signal and the state input pattern stored in the pattern database in an overlapping manner.

Images